MiniMax Open Source H3: Apart from Generosity, It's a Well Calculated Problem MiniMax has released the weights of the H3 video model. On the surface, it may seem like a good thing in the open source community, but companies always need to make profits. Now that the license has also changed, it's more like a well calculated question. The complete H3 is a three-stage assembly line. The front-end Context IR is responsible for understanding the relationship between the text, images, audio, and reference videos you feed in. The Regenerate-2K in the later stage regenerates the 768p result along with the original material back to 2K. The H3 Base in the middle stage has a parameter of 33B and can produce videos with 4 to 15 seconds, 768p, and stereo sound locally. MiniMax only open-source the middle section. The first and second parts are locked in the API. Interestingly, the official model card states that Context IR is crucial for the final quality, and it is strongly recommended to proceed with the assembly line. This creates a natural funnel: if you want a complete 2K effect with 768p running locally, go back to my API. Open source is a hook, API is a closed loop. The license agreement has a restricted threshold. Excluding the United States, the European Union, the United Kingdom, and South Korea. There are no flaws in compliance, the EU AI Act is being enforced, and copyright lawsuits in the United States are intensifying. But in business, this result is too coincidental: these four areas happen to be the markets with the strongest payment capabilities in the world, and 73% of MiniMax's revenue in 2025 comes from overseas. The free self deployment of weights is only available to places that do not affect the main revenue disk, and high-value customers stay in the API to receive payment. The income threshold is set to 20 million US dollars. Organizations below this line are free to use for commercial purposes, while those above it require separate authorization. This is a standard open source funnel: the free layer eats up long tail developers and small and medium-sized teams, turning large customers who can afford to pay into business negotiation partners. Lightricks' LTX-2 uses the same number of lanes, with a threshold of 10 million, which is similar to the gameplay of Kimi K3. The essence is clear: without distribution channels, the developer ecosystem is the only local monopoly that MiniMax can build. Open source is a choice forced by structure. Today, let's talk about the three most deadly things about the moat. Are assets still scarce. Once the weight is made public, the weight of 33B itself is no longer scarce. Scarcity is forced to shift to the two segments without open source and self built computing power. Among them, Context IR is the most worthy of scrutiny. The official reason for not being open source is that it relies on multi-stage workflows and multiple hosting services, making it difficult to package in engineering. This reason is valid, but when the model is broken down, it is divided into multiple sets of services, essentially a orchestration system, and the community is most convenient to reverse. The official also provided Prompting Guidance, which is equivalent to handing over a self built roadmap to you. There is a high probability that a community version replacement will appear within six months. Is copying expensive. There is an awkward reality here: there is already a more liberal option within the open source camp. Alibaba's Wan 2.7 achieved 1080 Elo on the audio and video editing chart, only behind H3's 1130 by 50 points. The license is Apache 2.0, with zero geographical restrictions and zero revenue threshold. Tencent's lightweight version can run on consumer grade graphics cards. The reason why developers choose H3 must be that the ability difference is so large that it is worth accepting the restrictions and the threshold of starting with four cards. Fifty points can be perceived in blind testing, but it cannot be considered crushing. Distribution loop. This is MiniMax's most fatal weakness. The video value chain consists of generation, distribution, and monetization. Bytes hold Tiktok and dream, Kwai holds the master station and Kering, and the model is just a loop in the closed loop. MiniMax has Conch AI and API in its hands. 236 million users and 214000 enterprise customers, these numbers indicate how many people use it, but they do not indicate that it has captured the channel. The videos generated by it will eventually be sent to Tiktok and YouTube, and the owners of these channels are also selling video models themselves. Dividing the market into four quadrants based on whether it has been distributed or not, MiniMax sits in the most difficult compartment: an open source+non distributed closed loop. This type of license requires both direct monetization of the model and open-source transformation of the ecosystem, and the license must be a limited community license. Bytes and Kwai sit in the strongest grid, and can be consumed with free and subsidies. Alibaba and Tencent are sitting in the forefront of open-source and factory support. Models do not expect to directly make money, and money is collected in the cloud. They do not need to use model revenue to support themselves. Price narratives should also be discounted. Officially, 2K per second is less than one-third of the mainstream. Using third-party verification with the same caliber, H3 is $7.80 per minute, Gemini Omni Flash is $6.00, and Seedance 720p is $5.57. In terms of absolute price, H3 is not the cheapest. The premise for the advantage to be established is consistent resolution: default to 2K, others default to 720p, which is only one-third of the unit pixel conversion. High resolution commercial delivery is a real advantage, while low resolution lightweight scenes have no moats. And that area happens to be the largest in terms of quantity, still controlled by opponents with distribution using a free strategy. Among the three cards, the underestimated one is the compatibility of domestic chips. In the early stages of H3 design, compatibility with multiple domestic chips was included, and one official motivation for opening the weight was to accelerate the adaptation of domestic chips. The business logic chain is clear: domestic computing power manufacturers need robust multimodal models to validate their own chips, domestic government and enterprises need compliant and controllable localized video generation, MiniMax needs to break away from a single computing power dependence and open up a market that ByteDance and Google cannot enter. This is not reflected in the Elo chart, but in the revenue structure. The domestic revenue share of MiniMax has increased from less than 30% to nearly 40%. H3 is a good product and a successful tactical move. Ranked first on the video editing chart, Wen Sheng Tu Sheng Video ranks second, with integrated sound and image, multimodal natural language instructions, and 2K regeneration. There is currently no opponent in the open weight model. But it did not solve MiniMax's strategic problem in the video arena. It buys developer minds, a ticket to domestic chips, and a good news cycle. It cannot buy distribution channels or pricing power. The product is good enough, but the business is not stable enough, but its valuation is only HKD 87 billion, much lower than the valuations of other video models, not to mention that minimax has other models. The biggest disadvantage is that it is not resource intensive enough, and each product is half baked